An author name disambiguation method and related apparatus

Abstract

The application discloses an author name disambiguation method and related device. In the method, author information to be analyzed and a candidate information set are acquired; multiple dimension similarity mining is respectively performed between multiple pieces of author information in the candidate information set and the author information to be analyzed; a cooperation network is one of the multiple dimensions. The multiple dimension similarity mining results between the same piece of author information in the candidate information set and the author information to be analyzed are fused to obtain multiple dimension similarity fusion results, and a matching result of the author to be analyzed and an author in the candidate information set is obtained according to the multiple dimension similarity fusion results. The scheme improves the effectiveness of the cooperation network in author name disambiguation, realizes deep understanding of structural topological characteristics and dynamic grasping of network evolution mode, has higher accuracy than the prior art, and can further improve the accuracy of author name disambiguation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an author name disambiguation method and related device. BACKGROUND

[0002] With the continuous progress of society and the rapid development of science and technology, the data volume of all kinds of literature on the network presents an explosive growth. The forms of massive literature are various, such as papers, patents, fund projects, software copyrights, monographs, etc. Some factors may cause the author information in the works information published by the same author to be different, so it is necessary to disambiguate the author name in order to accurately identify and match the different name variants of the same author appearing in different literature. There are many factors that cause the author information to be different, such as: change of the author's employment agency, same name of different authors, name spelling change, etc. At present, for the research direction of author name disambiguation, the main methods are rule-based string matching method and simple machine learning method. The existing methods usually only consider basic features such as name and organization, and the accuracy is generally between 70-85%, and the processing time is relatively long. Especially in the aspect of cooperation network analysis, the existing technology only stays at the level of simple co-author counting, and fails to deeply mine the network topology structure features and time sequence evolution mode, resulting in limited recognition ability of complex academic relations, and the accuracy of author name disambiguation still needs to be improved. SUMMARY

[0003] Based on the above problems, the present application provides an author name disambiguation method and related device, aiming to improve the accuracy of author name disambiguation.

[0004] The embodiments of the present application disclose the following technical solutions:

[0005] The first aspect of the present application provides an author name disambiguation method, which comprises:

[0006] obtaining a piece of author information to be analyzed and a candidate information set; the candidate information set comprises a plurality of pieces of author information; the author information to be analyzed comprises a co-author list;

[0007] respectively performing multi-dimensional similarity mining between the plurality of pieces of author information and the author information to be analyzed; wherein the cooperation network is one of the plurality of dimensions, the authors are represented by nodes in the cooperation network, and the cooperation relationship between nodes is represented by a connection between nodes; the similarity mining of the cooperation network comprises at least one of: direct cooperation relationship similarity mining based on the cooperation network, indirect network relationship similarity mining based on the cooperation network, network structure topology feature similarity mining based on the cooperation network, or time sequence network evolution similarity mining based on the cooperation network;

[0008] fuse results of the multi-dimension similarity mining between the same author information in the candidate information set and the author information to be analyzed to obtain multi-dimension similarity fusion results;

[0009] According to the multi-dimension similarity fusion results of the multi-piece author information and the author information to be analyzed, a matching result of the author to be analyzed and the author in the candidate information set is obtained.

[0010] The second aspect of the present application provides an author name disambiguation device, which comprises:

[0011] An information acquisition module is configured to acquire a piece of author information to be analyzed and a candidate information set; the candidate information set comprises multi-piece author information; the author information to be analyzed comprises a co-author list;

[0012] A multi-dimension similarity mining module is configured to respectively perform multi-dimension similarity mining between the multi-piece author information and the author information to be analyzed; wherein, a cooperation network is one of the multi-dimensions, an author is represented by a node in the cooperation network, and a cooperation relationship between nodes is represented by a connection line; the similarity mining of the cooperation network comprises at least one of the following: direct cooperation relationship similarity mining based on the cooperation network, indirect network relationship similarity mining based on the cooperation network, network structure topology feature similarity mining based on the cooperation network, or time sequence network evolution similarity mining based on the cooperation network;

[0013] A multi-dimension similarity fusion module is configured to fuse results of the multi-dimension similarity mining between the same author information in the candidate information set and the author information to be analyzed to obtain multi-dimension similarity fusion results;

[0014] A matching module is configured to obtain a matching result of the author to be analyzed and the author in the candidate information set according to the multi-dimension similarity fusion results of the multi-piece author information and the author information to be analyzed.

[0015] The third aspect of the present application provides an author name disambiguation device, which comprises a processor and a memory in mutual communication connection;

[0016] The memory stores a computer program;

[0017] The processor is configured to run the computer program to realize the author name disambiguation method as introduced in the first aspect.

[0018] The fourth aspect of the present application provides a computer readable storage medium, which stores a computer program; the computer program is configured to realize the steps of the author name disambiguation method as introduced in the first aspect when executed by a processor.

[0019] Compared with the prior art, the present application has the following beneficial effects:

[0020] In the technical solution of the present application, in order to realize accurate disambiguation of author names, first, a piece of author information to be analyzed and a candidate information set are obtained; the candidate information set includes multiple pieces of author information; the author information to be analyzed includes a co-author list. Then, multiple dimension similarity mining is respectively performed between the multiple pieces of author information and the author information to be analyzed; among them, the cooperation network is one of the multiple dimensions, and in the cooperation network, the authors are represented by nodes, and the cooperation relationship between the nodes is represented by the connection between the nodes. After that, the results of the multiple dimension similarity mining between the same piece of author information in the candidate information set and the author information to be analyzed are fused to obtain a multiple dimension similarity fusion result. Finally, according to the multiple dimension similarity fusion results of the multiple pieces of author information and the author information to be analyzed, a matching result of the author to be analyzed and the authors in the candidate information set is obtained.

[0021] In the technical solution of the present application, multiple dimension similarity mining of author information is considered, which involves mining of the cooperation network. And in the process of mining the cooperation network, not only the similarity of direct cooperation relationship and the similarity of indirect network relationship can be mined, but also the similarity of network structure topological features in the cooperation network and the similarity of time sequence network evolution are particularly proposed, which greatly improves the effectiveness of the cooperation network in author name disambiguation. For mining of the network structure topological feature similarity and the time sequence network evolution similarity, deep understanding of the structure topological features and dynamic grasp of the network evolution mode are realized, so that the multiple dimension similarity fusion result of the similarity mining result of the cooperation network dimension in the present application has higher accuracy than the prior art, and the accuracy of author name disambiguation can be further improved. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or the prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0023] Figure 1 A flowchart of an author name disambiguation method provided by an embodiment of the present application;

[0024] Figure 2 An implementation architecture diagram of the author name disambiguation method;

[0025] Figure 3 A schematic diagram of four aspects of cooperation network similarity mining proposed by an embodiment of the present application;

[0026] Figure 4 Fig. 6 is a schematic diagram of six-dimension similarity mining of author information and obtaining of a corresponding similarity mining result;

[0027] Figure 5 Fig. 7 is an implementation architecture diagram of an author name disambiguation method provided by an embodiment of the present application;

[0028] Figure 6 Fig. 8 is a structural schematic diagram of an author name disambiguation apparatus provided by an embodiment of the present application. DETAILED DESCRIPTION

[0029] At present, in the research direction of author name disambiguation, there is research on cooperation networks, but generally only stays at the level of simple co-author counting, and the research on cooperation networks is not sufficient. Therefore, the prior art fails to mine the knowledge content directly reflected or implicitly reflected in the cooperation network to the greatest extent, resulting in very limited recognition ability of complex academic relationships, and further affecting the effectiveness of the cooperation network on author name disambiguation.

[0030] The inventor proposes an author name disambiguation method and related apparatus. In the technical solution of the present application, the similarity mining of multiple dimensions of author information is focused on, which also includes the similarity mining of cooperation networks. The similarity mining of cooperation networks can be started from multiple aspects, for example, the similarity mining of direct cooperation relationship, the similarity mining of indirect network relationship, the similarity mining of network structure topology characteristics, and the similarity mining of time sequence network evolution. Through the similarity mining of one or more of the above aspects of the cooperation network, and the fusion of the results of the similarity mining of other dimensions, the multiple-dimension similarity fusion results between the same author information in the candidate information set and the author information to be analyzed can be obtained. Finally, according to the multiple-dimension similarity fusion results of the multiple pieces of author information and the author information to be analyzed respectively, the matching result of the author to be analyzed and the author in the candidate information set is determined. The matching result reflects the disambiguation of the author name of the author to be analyzed.

[0031] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0032] It should be noted that before the author name disambiguation method is formally implemented, the embodiments of the present application can use intelligent two-level adaptive caching to initialize the system. For example, create a memory LRU cache (default 3000 items), set the adaptive TTL mechanism (adjust dynamically according to data access frequency for 7-30 days); establish a SQLite persistent cache, use WAL mode to optimize concurrent access, set 256MB memory mapping; implement cache hit rate statistics and performance monitoring, and automatically expand when the hit rate is less than 70%. The two-level adaptive cache is an intelligent caching system combining memory LRU cache and SQLite persistent cache, which can automatically adjust the caching strategy according to the access mode and data characteristics, and supports separate management of query result caching and calculation result caching.

[0033] Referring to Figure 1 , the figure is a flowchart of an author name disambiguation method provided by the embodiments of the present application. As Figure 1 indicated, the method comprises:

[0034] S101, obtaining a piece of author information to be analyzed and a candidate information set.

[0035] In the present application, in order to realize the disambiguation of the author name, a piece of author information to be analyzed and a candidate information set are first obtained. The candidate information set includes multiple pieces of author information. Figure 2 For the implementation architecture of the author name disambiguation method, refer to Figure 2 for an example. The candidate information set contains N pieces of author information, which can be referred to as “author information entry 1”, “author information entry 2”, …, and “author information entry N” for easy distinction.

[0036] An example implementation manner of obtaining the candidate information set is proposed in the present application. Before obtaining the candidate information set, first, the author to be analyzed is determined as the author to be disambiguated in this time based on a piece of author information to be analyzed. Based on the obtained author information to be analyzed, the manner of obtaining the candidate information set can include:

[0037] In the author information library, based on the surname and the first letter of the author information to be analyzed, author information matching is performed to obtain a first matching information set meeting the matching condition; in the first matching information set, based on the institution keyword and the year range of the author information to be analyzed, author information matching is performed to obtain a second matching information set meeting the matching condition; in the second matching information set, based on the identity recognition code of the author information to be analyzed, author information matching is performed to obtain a third matching information set meeting the matching condition as the candidate information set. It can be seen that in the embodiments of the present application, the candidate information set is obtained through a three-level screening strategy based on the surname and the first letter, the institution keyword and the year range, and the identity recognition code.

[0038] It can be known from the above description that, in order to screen out a candidate information set for similarity mining and matching with author information to be analyzed, the application can adopt a three-level screening manner, the first level performs one round of screening from the surname and the first letter, the second level performs another round of screening from the institution keyword and the year range, and the third level performs still another round of screening from the identity code. The candidate information set finally constructed through the three-level screening has a high matching degree with the author information to be analyzed in terms of the name, the first letter, the institution, the year range, and the identity code, and is beneficial to realize rapid positioning of author information with a high correlation degree, and reduce the candidate information set to a small amount of data. In an optional example, the first-level screening controls the number of information entries in the first matching information set to be within 500 entries, and the second-level screening controls the number of information entries in the second matching information set to be within 200 entries.

[0039] In actual application, before the author information meeting the matching condition is filtered from the author information library and added to the candidate information set for the author information to be analyzed, the technical solution of the application can also perform data quality evaluation on the author information to be analyzed. For example, the author information to be analyzed can be subjected to integrity checking, format verification, and outlier detection. The received author information to be analyzed at least includes a co-author list, and can further include a standardized name, an institution list, a publication year, and an identity code. The data quality evaluation manner can be different for different dimensions of information in the above information, and can be specifically operated according to the designed integrity checking manner, format verification manner, and outlier detection manner, which is not limited here.

[0040] After the above data quality evaluation is performed, a query fingerprint can be established for each piece of author information for caching index, and the specific form of the query fingerprint can be MD5 (name + institution keyword + year range). Wherein, MD5 represents MD5 Message-Digest Algorithm, a widely used cryptographic hash function, which can generate a 128-bit (16-byte) hash value. Establishing the query fingerprint for each piece of author information and caching the index are helpful to index to the condition-matched author information when the candidate information set needs to be screened out, and realize efficient establishment of the candidate information set.

[0041] In order to intelligently construct the candidate information set and optimize the query of author information, a query plan generator can be constructed, which automatically selects the optimal query path according to data distribution; or a batch query strategy is adopted to avoid the large impact of large data volume query results on memory. In addition, a query performance monitoring mechanism can be established, and when the monitoring query time exceeds the threshold, a backup query strategy is automatically switched to avoid long waiting time before disambiguation and improve the efficiency of author name disambiguation.

[0042] In the above description of the query plan generator, the "data distribution" refers to the name frequency distribution in the author database (such as the number of occurrences of high-frequency surnames such as Smith and Wang), institutional distribution characteristics (the number of authors of each institution, the proportion of national distribution), ID coverage distribution (the coverage proportion of ORCID and Scopus ID), and time distribution characteristics (the activity of authors in different year segments). Statistical information such as. And "query path" refers to different query strategies designed according to these distribution characteristics, such as:

[0043] When the input author has a reliable ID and the ID coverage is high, select the "ID priority path" (first use ID exact match and then verify name and institution);

[0044] When a high-frequency surname is encountered, select the "name-institution joint path" (filter with name and institution keywords at the same time to narrow down the candidate set);

[0045] When the collaborator information is rich, select the "collaboration network priority path" (first build a candidate set through the collaboration network and then verify the name and institution).

[0046] The query plan generator analyzes the input author information (whether there is an ID, whether the surname is common, whether the collaborator information is rich, etc.), combines data distribution statistics and historical query performance, automatically calculates the expected efficiency score of each path, and then selects the optimal path to execute the query, and dynamically adjusts the strategy according to the actual query effect, thereby realizing intelligent query optimization and load balancing.

[0047] S102, respectively, multiple dimension similarity mining between multiple author information and author information to be analyzed.

[0048] In this application, each piece of author information in the candidate information set needs to be mined for similarity in multiple dimensions with the aforementioned author information to be analyzed.

[0049] Among the plurality of dimensions, a dimension of a cooperation network is included. The cooperation network is constructed as an undirected weighted graph G=(V, E, W), where V represents a set of author nodes, E represents a set of cooperation relationship edges, and W represents a set of edge weights; the edge weight is calculated based on cooperation frequency and a time decay factor: w(i, j) = count(i, j) x exp(-l x (current year - cooperation year)), where i and j represent two author nodes in the graph, and l = 0.1 / year. The cooperation network is actually a complete graph network structure, and can also be understood as a global cooperation network, rather than a radial network with a single author as the core. Specifically, the global cooperation network is a global graph containing all author nodes, and there is an edge between any two authors with a cooperation relationship, and the weight of the edge represents the cooperation strength. When analyzing a specific author, a self-network of the author can be extracted, which can also be understood as a local network or a sub-network. In the self-network, the author is at the center in a radial manner, and the self-network includes the author's first-degree neighbors (direct collaborators), second-degree neighbors, third-degree neighbors, and the like.

[0050] In the similarity mining between two authors with respect to the cooperation network dimension, the network topological structure features (such as local clustering coefficient, structural hole position, network centrality) and time sequence network evolution of the self-networks of the two authors are compared, rather than simple radial matching.

[0051] That is, in the embodiments of the present application, the global cooperation network is constructed based on the collaborator list, and when analyzing a specific author, a self-network centered on the author is extracted for network feature analysis.

[0052] In the example implementation, the construction of the cooperation network has the following points:

[0053] The cooperation relationship is extracted, and a weighted directed graph is constructed. The weight in the weighted directed graph can be calculated based on the product of the cooperation frequency and the time decay factor. In addition, network structure topological features such as node degree centrality, clustering coefficient, betweenness centrality, and shortest path length can be calculated. In addition, an author influence score model can be constructed, and the score Score = a x degree centrality + b x clustering coefficient + g x article quality.

[0054] Figure 3 The four aspects of the cooperation network similarity mining proposed in the embodiments of the present application are shown in the schematic diagram. Figure 3 As shown, with respect to the cooperation network dimension, the embodiments of the present application propose to perform similarity mining in at least one of the following four aspects: direct cooperation relationship similarity mining based on the cooperation network, indirect network relationship similarity mining based on the cooperation network, network structure topological feature similarity mining based on the cooperation network, or time sequence network evolution similarity mining based on the cooperation network.

[0055] Among them, the direct cooperation relationship similarity mining focuses on the similarity of the cooperation relationship directly reflected by the respective cooperation networks of the two authors compared with each other;

[0056] The indirect network relationship similarity mining focuses on the potential cooperation relationship or academic circle connection reflected by the respective cooperation networks of the two authors compared with each other;

[0057] The network structure topology feature similarity mining focuses on the similarity of the structure topology features of the respective cooperation networks of the two authors compared with each other, and analyzes the similarity of the authors through the deep analysis of the network structure;

[0058] The time sequence network evolution similarity mining focuses on the similarity of the dynamic information over time reflected by the respective cooperation networks of the two authors compared with each other.

[0059] The above has briefly described the similarity mining of the cooperation network dimension, and the specific mining methods of the above four aspects will be described in detail below, and no more description is made here.

[0060] In addition to the cooperation network, the step S102 performs similarity mining in multiple dimensions, and the multiple dimensions can further include an identity code, a name, an organization, a time and an alias. Figure 4 An example of six-dimension similarity mining of author information and obtaining the corresponding similarity mining results is shown in the figure. The similarity mining of these dimensions can be performed based on the specific information types contained in the multiple pieces of author information in the candidate information set and the information types of the author to be analyzed contained in the author information to be analyzed, or further processing is performed on the basis of the specific information types contained in the author information to complete the similarity mining work. Through S102, the identity code similarity mining result between each piece of author information in the candidate information set and the author information to be analyzed, the name similarity mining result, the organization similarity mining result, the cooperation network similarity mining result, the time similarity mining result and the alias similarity mining result can be obtained. The similarity mining results of these different dimensions will be fused according to the correspondence with the author information in the candidate information set in the following step S103.

[0061] S103, the results of the multiple-dimension similarity mining between the same author information in the candidate information set and the author information to be analyzed are fused to obtain multiple-dimension similarity fusion results.

[0062] In the optional implementation, in combination with the above six dimensions, the example of similarity mining can be the result of identity code similarity mining between the same author information in the candidate information set and the author information to be analyzed, the result of name similarity mining, the result of organization similarity mining, the result of cooperation network similarity mining, the result of time similarity mining and the result of alias similarity mining. The multiple dimension similarity fusion results between the same author information and the author information to be analyzed are obtained by fusing the results according to the weight of the corresponding dimension.

[0063] For the result of identity code similarity mining, the result of name similarity mining, the result of organization similarity mining, the result of cooperation network similarity mining, the result of time similarity mining and the result of alias similarity mining, each can be represented by a numerical value, so that the fusion can be performed by weighting, and finally the multiple dimension similarity fusion results are obtained.

[0064] For example, the weight of the identity code dimension is 0.25, the weight of the name dimension is 0.4, the weight of the organization dimension is 0.15, the weight of the cooperation network dimension is 0.15, the weight of the time dimension is 0.03, and the weight of the alias dimension is 0.02. It should be noted that the weight of each dimension can be set or adjusted according to the needs in combination with different fusion needs or the focus or secondary consideration of the similarity of a specific dimension. The above is only an example of weight setting, and is not limited to the specific weight value.

[0065] S104, according to the multiple dimension similarity fusion results of the multiple author information and the author information to be analyzed, obtaining the matching result of the author to be analyzed and the author in the candidate information set.

[0066] Each piece of author information in the candidate information set can be subjected to multiple dimension similarity mining with the author information to be analyzed, and multiple dimension similarity fusion results are obtained, such as "multiple dimension similarity fusion result 1", "multiple dimension similarity fusion result 2", …, "multiple dimension similarity fusion result N" shown in the above. Figure 2 After obtaining these multiple dimension similarity fusion results, the matching result of the author to be analyzed and the author in the candidate information set can be obtained by comparing their numerical values. For example, the author information with the highest multiple dimension similarity fusion result value with the author information to be analyzed is determined as the matching author information, and it is considered that the author information and the author information to be analyzed substantially correspond to the same author, thereby completing the disambiguation of the author name.

[0067] In the technical solution of the present application, multiple dimensions of similarity of author information are considered, which involves mining of a cooperation network. In the process of mining the cooperation network, not only the similarity of direct cooperation relationship and the similarity of indirect network relationship can be mined, but also the similarity of network structure topology features in the cooperation network and the similarity of time sequence network evolution are proposed, which greatly improves the effectiveness of the cooperation network in author name disambiguation. For mining of the network structure topology feature similarity and the time sequence network evolution similarity, a deep understanding of the structure topology features and a dynamic grasp of the network evolution pattern are realized, so that the multiple dimension similarity fusion results of the similarity mining results in the cooperation network dimension in the present application have higher accuracy than the prior art, and the accuracy of author name disambiguation can be further improved.

[0068] It is mentioned above that in the cooperation network dimension, at least one of the following four aspects of similarity mining can be performed: direct cooperation relationship similarity mining based on the cooperation network, indirect network relationship similarity mining based on the cooperation network, network structure topology feature similarity mining based on the cooperation network, or time sequence network evolution similarity mining based on the cooperation network. The similarity mining processes of the four aspects will be introduced respectively below. For convenience of introduction, the author to which a certain author information in the candidate information set belongs is called the target author, the local network extracted from the global cooperation network for the author to be analyzed is called the first cooperation network, and the local network extracted from the global cooperation network for the target author is called the second cooperation network, so that the similarity mining processes of the cooperation network in each aspect are described more concisely and accurately.

[0069] (1) Network structure topology feature similarity mining.

[0070] In the embodiment of the present application, the network structure topology feature similarity mining of the target author information in the candidate information set and the author information to be analyzed based on the cooperation network includes:

[0071] The network structure topology features of the first cooperation network are analyzed to obtain the local clustering coefficient, the structural hole position and the network centrality of the first cooperation network, and the network structure topology features of the second cooperation network are analyzed to obtain the local clustering coefficient, the structural hole position and the network centrality of the second cooperation network;

[0072] Based on the local clustering coefficient of the first cooperation network and the local clustering coefficient of the second cooperation network, the local clustering similarity between the target author and the author to be analyzed is calculated; based on the structural hole position of the first cooperation network and the structural hole position of the second cooperation network, the structural hole position similarity between the target author and the author to be analyzed is calculated; based on the network centrality of the first cooperation network and the network centrality of the second cooperation network, the network centrality similarity between the target author and the author to be analyzed is calculated.

[0073] Finally, according to the local clustering similarity, the structural hole position similarity and the network centrality similarity, the network structure topology feature similarity mining result of the target author and the author to be analyzed is calculated.

[0074] As an example, the local clustering similarity, the structural hole position similarity and the network centrality similarity each have a fusion weight, the local clustering similarity, the structural hole position similarity and the network centrality similarity are weighted and summed based on the fusion weights of the three, and the final calculation result is taken as the network structure topology feature similarity mining result of the target author and the author to be analyzed.

[0075] Among them, the local clustering coefficient is used to measure the closeness of the network where the author is located; the structural hole position is used to identify the bridge role of the author in the network; and the network centrality is used to realize the comprehensive evaluation of the degree centrality and the betweenness centrality.

[0076] The local clustering coefficient calculation formula is: C(i) = 2×ei / (ki×(ki-1)), where ei is the actual edge number between the neighbors of node i, and ki is the degree of node i.

[0077] The structural hole position is calculated by effective scale:

[0078]

[0079] In the formula, p iq represents the standardized connection strength of node i and node q, p qj represents the standardized connection strength of node q and node j, and the sum is the sum of all nodes q except node i and node j. The larger the effective scale value ES(i), the more important the structural hole position of node i in the network, that is, the neighbors of node i lack direct connection, and node i plays an important bridge role. In the formula, is the sum of all neighbor nodes j of node i, and is the "uniqueness" index corresponding to each neighbor node j, that is, the degree to which node j depends on i without passing through other nodes. By adding the "uniqueness" index corresponding to each neighbor node j, the effective scale value of node i is obtained.

[0080] The network centrality adopts a weighted combination of degree centrality and betweenness centrality: NC(i) = a x DC(i) + β x BC(i), where a + β = 1, DC represents the degree centrality, and BC represents the betweenness centrality.

[0081] In the embodiments of the present application, based on the network structure topology feature analysis driven by the graph neural network, including the degree centrality, the clustering coefficient, the structural hole analysis and other advanced network features, the technical gap of the cooperation network deep analysis is filled.

[0082] (2) Time sequence network evolution similarity mining.

[0083] In the embodiments of the present application, the target author information in the candidate information set and the author information to be analyzed are subjected to time sequence network evolution similarity mining based on the cooperation network, including:

[0084] The first cooperation network is subjected to time window analysis to obtain a first time window analysis result, and a dynamic change mode of the first cooperation network is identified; the second cooperation network is subjected to time window analysis to obtain a second time window analysis result, and a dynamic change mode of the second cooperation network is identified;

[0085] Based on the first time window analysis result and the second time window analysis result, a network time window similarity of the target author and the author to be analyzed is obtained;

[0086] Based on the dynamic change mode of the first cooperation network and the dynamic change mode of the second cooperation network, a network dynamic change mode similarity of the target author and the author to be analyzed is obtained;

[0087] According to the network time window similarity and the network dynamic change mode similarity, a time sequence network evolution similarity mining result of the target author and the author to be analyzed is calculated.

[0088] As an example, the network time window similarity and the network dynamic change mode similarity each have a fusion weight, the network time window similarity and the network dynamic change mode similarity are weighted and summed based on the fusion weights of the two, and the finally calculated result is taken as the time sequence network evolution similarity mining result of the target author and the author to be analyzed.

[0089] It can be understood that the cooperator list in the author information not only reflects the object of specific cooperation, but also has a time attribute, for example, cooperating to publish a work in July 2021 and cooperating to publish another work in November 2022. Therefore, through the time window analysis of the cooperation network, the dynamic characteristics of the cooperation relationship can be identified, and then the similarity of the network time window and the similarity of the network dynamic change pattern are mined. It can be understood that if it is the same author, the network time window and the network dynamic change pattern should tend to be similar. In this application, through the above-mentioned time sequence network evolution similarity mining, compared with the existing technology for mining the cooperation network, the potential dynamic change attribute of the network can be more accurately captured, which is used as an aspect of cooperation network similarity mining. The accuracy and effectiveness of author name disambiguation of the cooperation network can be improved.

[0090] (3) Direct cooperation relationship similarity mining.

[0091] In the embodiment of the application, the target author information in the candidate information set and the author information to be analyzed are subjected to direct cooperation relationship similarity mining based on the cooperation network, which includes:

[0092] According to the first cooperation network, the first cooperation strength and the first cooperation density of each author in the cooperator list are calculated; according to the second cooperation network, the second cooperation strength and the second cooperation density of each author in the cooperator list are calculated;

[0093] According to the first cooperation strength and the second cooperation strength of the same author in the cooperator list between the author to be analyzed and the target author, the cooperation strength similarity of the target author and the author to be analyzed is obtained;

[0094] According to the first cooperation density and the second cooperation density of the same author in the cooperator list between the author to be analyzed and the target author, the cooperation density similarity of the target author and the author to be analyzed is obtained;

[0095] According to the cooperation strength similarity and the cooperation density similarity, the direct cooperation relationship similarity mining result of the target author and the author to be analyzed is calculated.

[0096] As an example, the cooperation strength similarity and the cooperation density similarity each have a fusion weight, and the cooperation strength similarity and the cooperation density similarity are weighted and summed based on the fusion weight of the two, and the final calculation result is taken as the direct cooperation relationship similarity mining result of the target author and the author to be analyzed.

[0097] The following provides an example of calculating cooperation strength and cooperation density:

[0098] Cooperation strength = number of cooperation times x time decay factor x literature quality weight;

[0099] Cooperation density = Number of co-publications / max [Author A's publication number, Author B's publication number].

[0100] In the formula of cooperation density, Author A and Author B refer to the two authors used to calculate the cooperation density.

[0101] As can be seen from the above formula, the cooperation intensity is positively correlated with the number of cooperation times, the time decay factor and the literature quality weight.

[0102] The more distant the cooperation, the smaller the time decay factor; the more recent the cooperation, the larger the time decay factor. The reason for such design is that the recent cooperation relationship can better reflect the current academic contact and active state of the author, and has higher reference value in author disambiguation, while the distant cooperation relationship may no longer be active or have lower relevance. A typical time decay function is implemented as follows:

[0103] decay_factor = exp(-λ × (current year - cooperation year));

[0104] Wherein, decay_factor represents the time decay factor, λ represents the decay parameter, and the value λ=0.1 / year is taken, so that the time decay factor of the cooperation relationship 5 years ago is changed to exp(-0.5)≈0.61, i.e. about 60% of the weight, and the time decay factor of the cooperation relationship 10 years ago is changed to about 0.37. When the cooperation occurs in the current year, the time decay factor is 1.0 (the maximum value), and as the time difference increases, the time decay factor gradually decreases and tends to 0. Thus, more weight is given to the recent cooperation and less weight is given to the distant cooperation in the cooperation intensity calculation.

[0105] The cooperation density is positively correlated with the number of co-publications and negatively correlated with the maximum value of the number of publications of the two authors.

[0106] In this application, considering the direct cooperation relationship reflected by the cooperation intensity and the cooperation density, by comparing the similarity of such direct cooperation relationship, the commonality of the cooperation relationship between the target author and the author to be analyzed can be accurately grasped, and then the same author can be accurately identified to realize author name disambiguation.

[0107] (4) Indirect network relationship similarity mining.

[0108] In the embodiment of the present application, the target author information in the candidate information set and the author information to be analyzed are subjected to indirect network relationship similarity mining based on the cooperation network, which includes:

[0109] The two-degree connection nodes and three-degree connection nodes of the author to be analyzed are identified from the first cooperation network; the two-degree connection nodes and three-degree connection nodes of the target author are identified from the second cooperation network.

[0110] obtaining a two-degree connection similarity between the target author and the author to be analyzed based on the two-degree connection node of the author to be analyzed and the two-degree connection node of the target author, and obtaining a three-degree connection similarity between the target author and the author to be analyzed based on the three-degree connection node of the author to be analyzed and the three-degree connection node of the target author;

[0111] According to the two-degree connection similarity and the three-degree connection similarity, an indirect network relationship similarity mining result between the target author and the author to be analyzed is calculated.

[0112] The two-degree connection node of the author to be analyzed in the first cooperation network can be constructed based on the cooperator list of the one-degree connection node of the author to be analyzed, and the three-degree connection node of the author to be analyzed can be constructed based on the cooperator list of the two-degree connection node of the author to be analyzed. Similarly, the two-degree connection node of the target author in the second cooperation network can be constructed based on the cooperator list of the one-degree connection node of the target author, and the three-degree connection node of the target author can be constructed based on the cooperator list of the two-degree connection node of the target author.

[0113] As an example, the two-degree connection similarity and the three-degree connection similarity each have a fusion weight, and the two-degree connection similarity and the three-degree connection similarity are weighted and summed based on the fusion weights of the two, and the final calculation result is taken as the indirect network relationship similarity mining result between the target author and the author to be analyzed.

[0114] The two-degree connection similarity can reflect the connection strength of the common cooperator, and the three-degree connection similarity can be used to identify the potential academic circle relationship. Therefore, the indirect network relationship similarity mining result calculated based on the two-degree connection similarity and the three-degree connection similarity effectively analyzes the commonality of the author to be analyzed and the target author in the indirect network relationship, deepens and extends the shallow understanding of the cooperation network, and thus helps to accurately identify the same author.

[0115] After the above four aspects of similarity mining based on the cooperation network are completed, the network structure topology feature similarity mining result, the time sequence network evolution similarity mining result, the direct cooperation relationship similarity mining result, and the indirect network relationship similarity mining result can be fused to obtain the cooperation network similarity mining result. For example, the similarity mining results of the above four aspects each have a fusion weight, and the four aspects of similarity mining results are weighted and summed using the corresponding fusion weights, and the final obtained result is taken as the cooperation network similarity mining result. The cooperation network similarity mining result and the similarity mining results of other dimensions can be fused with the results of steps S103 and several other dimensions similarity mining in the above embodiment method to obtain a multi-dimension similarity fusion result.

[0116] Having described in detail the similarity mining implementation of the cooperation network dimension, the similarity mining of the other five dimensions will be described below. Figure 4

[0117] (1) Institution-based similarity mining.

[0118] In this application, an institution level recognition system and a country code standardization mapping are pre-established, so that the diversification of the expression of the institution name and the hierarchical relationship can be effectively processed.

[0119] In the embodiment of the application, the institution-based similarity mining is performed on the target author information in the candidate information set and the author information to be analyzed, including:

[0120] The institution information in the target author information and the institution information in the author information to be analyzed are respectively subjected to standardization preprocessing; the standardization preprocessing includes removing the prefix of the institution type, extracting the core keywords of the institution, and applying the country code for standardization mapping. The removal of the prefix of the institution type can be the removal of (Department of / School of, etc.). The way of extracting the core keywords of the institution can be to remove common words such as "university" and "college". The application of the country code for standardization mapping can accurately match the corresponding country and geographical location.

[0121] The keywords of the two sets of institution information after standardization preprocessing are calculated for Jaccard similarity and fuzzy matching similarity; the institution preliminary similarity is calculated based on the Jaccard similarity and the fuzzy matching similarity; the institution preliminary similarity is optimized based on the matching of the geographical location of the two sets of institution information and the matching of the institution level, and the optimized similarity result is taken as the result of the institution similarity mining of the target author and the author to be analyzed.

[0122] As introduced above, the institution-based similarity mining process mainly includes two stages, which can also be understood as adopting a hierarchical matching strategy: first, the weighted average of the Jaccard similarity (the proportion of the intersection of the keywords of the two sets of institution information to the union, |A∩B| / |A∪B|, A∩B represents the intersection of the keywords of the two sets of institution information, and A∪B represents the union of the keywords of the two sets of institution information) and the fuzzy matching similarity (token matching of string based on edit distance) is obtained to obtain the institution preliminary similarity; and then the geographical and institution level optimization is performed.

[0123] In the example implementation, the optimization based on the geographical location and the institution level can include:

[0124] The geographical location optimization adds a geographical adjustment coefficient according to the country matching condition. For example, the same country +0.1~0.2, different countries but with international cooperation history +0.05, and completely different countries -0.1.​

[0125] The institutional level optimization adds an institutional level adjustment coefficient based on the level matching. For example, the same level is, for example, university + 0.1, and the cross-level is, for example, university vs. department - 0.05~0.15.

[0126] The calculation formula of the optimized similarity result can be expressed as:

[0127] Optimized similarity = min[1.0, preliminary similarity x (1+geographical adjustment coefficient+institutional level adjustment coefficient)];

[0128] For the similarity mining of institutions, through the above hierarchical matching strategy, both the text similarity basis and the realistic constraints of the geographical distribution and organizational structure of institutions are ensured, so that the obtained institution similarity mining result has higher rationality in numerical value, and thus accurate author name disambiguation is effectively realized.

[0129] The embodiments of the present application can also realize the clustering of institutional aliases through an algorithm, so as to automatically identify different expressions of the same institution.

[0130] The above-mentioned concepts of institutional level and standardized mapping of country code are mentioned. In order to more accurately realize the similarity mining of the institution dimension, the present application proposes to construct an institutional level identification system, a standardized mapping table of country code, and a feature library of geographical location and subject distribution of institutions. The following will be described.

[0131] Constructing the institutional level identification system: in the institutional level identification system, the level of a university is higher than that of an institute, and the level of an institute is higher than that of a department. For example, University (3) > Institute (2) > Department (1). Wherein University represents a university, Institute represents an institute, Department represents a department, and the number in the bracket represents the specific level of the institution in the institutional level identification system. The larger the number is, the higher the level is.

[0132] Constructing the standardized mapping table of country code: the standardized mapping table of country code supports more than 200 countries or regions in various forms of expression. Through the established standardized mapping table of country code, the expression of countries and regions can be accurately and unambiguously identified, the corresponding mapping of the expression of countries and regions is realized, and the geographical information of countries and regions is accurately determined.

[0133] The construction institution geographical location and subject distribution feature library: as the name implies, the library can assist in determining the geographical location of the institution. As for the subject distribution features in the library, although they are not applied in the institution similarity mining, they can be reserved for expansion or applied in other unexplained scenarios related to academic relationships or subject relationships. With the help of the distribution of subjects, the author's name can be accurately disambiguated.

[0134] The application of the country code for standardized mapping is as follows: based on the country code standardized mapping table, the country code in the institution information is standardized mapped;

[0135] The aforementioned matching of the geographical location of the two sets of institution information is based on the analysis and determination of the institution geographical location and subject distribution feature library; the aforementioned matching of the institution level of the two sets of institution information is based on the analysis and determination of the institution level identification system.

[0136] (2) Similarity mining based on name.

[0137] In the embodiments of the present application, the algorithm process of deep name similarity calculation is innovatively adopted, and specifically, the similarity mining based on name is performed on the target author information in the candidate information set and the author information to be analyzed, including:

[0138] The name information in the target author information and the name information in the author information to be analyzed are respectively subjected to Unicode standardization processing;

[0139] The two sets of name information subjected to Unicode standardization processing are subjected to compound surname recognition;

[0140] Based on the complete matching weight of the surname, the fuzzy matching weight of the surname, the full name token sorting matching weight, and the initial letter sequence matching weight, a multi-level matching result of the two sets of name information is calculated;

[0141] The cultural background matching is adopted for the two sets of name information; wherein the cultural background matching includes: for Chinese names, the pinyin conversion and initial letter matching are adopted;

[0142] According to the multi-level matching result and the result of the cultural background matching, the result of the similarity mining of the name of the target author and the author to be analyzed is generated.

[0143] The Unicode standardization processing can include NFD decomposition, ASCII conversion, and special character cleaning, etc. NFD (Normalization Form Decomposition) is a character normalization specification in the Unicode standard, which decomposes characters into basic graphemes or combined characters. This decomposition is based on the specification decomposition rules of Unicode, ensuring the compatibility of character encoding between different platforms.

[0144] The recognition of the compound surname can be detecting more than 100 preset compound surname prefixes such as van / von / de / della.

[0145] The embodiments of the present application propose multi-level matching calculation for name similarity mining, which involves complete matching of surnames, fuzzy matching of surnames, full name string (Token) sorting matching, and initial letter sequence matching. As an example, the complete matching weight of the surname is 0.4, the fuzzy matching weight of the surname is 0.3, the full name string (Token) sorting matching weight is 0.2, and the initial letter sequence matching weight is 0.1. That is, the weights of the above four levels of matching calculation decrease in turn. Through multi-level matching, the problem of poor accuracy caused by single-level name similarity mining is avoided.

[0146] In addition, cultural background adaptation is also performed for the two sets of name information, which reduces the difficulty of name similarity mining and assists in improving the accuracy of name matching between the target author and the author to be analyzed, in combination with related knowledge of cultural background.

[0147] (3) Time-based similarity mining.

[0148] In the embodiments of the present application, when similarity mining is performed for the time dimension, the similarity of the publication year of the works published by the author is considered, the overlap of the career track and the active period is analyzed, and the frequency similarity of the published works is analyzed. Specifically, in the embodiments of the present application, the target author information in the candidate information set and the author information to be analyzed are subjected to time-based similarity mining, including:

[0149] According to the publication year in the target author information and the publication year in the author information to be analyzed, the publication year similarity is calculated by using a Gaussian decay function;

[0150] According to the first publication year in the target author information, the academic career stage of the target author is inferred, and according to the first publication year in the author information to be analyzed, the academic career stage of the author to be analyzed is inferred. Based on the two academic career stages, career track analysis is performed to obtain a career track similarity analysis result;

[0151] According to the publication year in the target author information and the publication year in the author information to be analyzed, the overlap degree of the active period of the work publication is calculated;

[0152] According to the annual publication quantity in the target author information and the annual publication quantity in the author information to be analyzed, the publication frequency similarity of the target author and the author to be analyzed is calculated;

[0153] According to the publication year similarity, the career trajectory similarity analysis result, the overlap degree of the active period of the work publication, and the publication frequency similarity, the result of the time similarity mining of the target author and the author to be analyzed is generated.

[0154] In practical applications, the above publication year similarity, career trajectory similarity analysis result, overlap degree of the active period of the work publication, and publication frequency similarity each have a corresponding fusion weight. The similarity calculation results of the above four aspects can be weighted and summed according to the corresponding fusion weight, and the final result is used as the similarity mining result of the target author and the author to be analyzed in the time dimension. In the technical solution of the present application, through similarity mining in multiple aspects in the time dimension, not only the similarity of the publication-related explicit features (such as publication year and publication frequency) is analyzed, but also the similarity of the features with a certain time span and mode performance such as the overlap degree of the active period of the work publication is deeply extended, realizing the deep understanding and analysis of the commonality in the time dimension, and further helping to improve the effect of author name disambiguation.

[0155] (4) Similarity mining based on alias.

[0156] In the embodiments of the present application, similarity mining based on alias is performed on the target author information in the candidate information set and the author information to be analyzed, including:

[0157] Based on the author alias knowledge base, an alias is generated for the name information in the target author information, and based on the author alias knowledge base, an alias is generated for the name information in the author information to be analyzed;

[0158] Based on the generated two groups of aliases, an extended matching is performed to obtain an alias matching result;

[0159] Based on the usage frequency, time distribution, and source credibility, a confidence evaluation is performed on the alias matching result, and based on the evaluated confidence and the alias matching result, an alias similarity mining result of the target author and the author to be analyzed is generated.

[0160] The established author alias knowledge base can include historical aliases, common variants, misspellings, etc. The alias generation algorithm can be used to generate aliases for the name information in the author information to be analyzed. In the actual application, the frequency, time distribution, and source credibility can be used to evaluate the confidence of the alias matching results.

[0161] In addition, after generating the alias, before performing the alias matching, the confidence of the alias can be evaluated based on the frequency of use, the time distribution, and the source credibility. After the confidence of the alias is evaluated, the alias with a higher evaluation confidence is promoted to the alias matching process. The matching result of the alias is directly used as the alias similarity mining result.

[0162] The alias is used to expand the matching of the author information compared with each other, which expands the utilization rate of the name information in the author name disambiguation scene, increases the matching dimension, and improves the realizability of the author name disambiguation.

[0163] (5) Similarity mining based on identity recognition code.

[0164] In the embodiments of the present application, the target author information in the candidate information set and the author information to be analyzed are subjected to similarity mining based on the identity recognition code, which includes:

[0165] The ORCID in the target author information and the ORCID in the author information to be analyzed are subjected to complete matching to obtain a first matching score;

[0166] The Scopus ID in the target author information and the Scopus ID in the author information to be analyzed are subjected to fuzzy matching using the edit distance algorithm to obtain a second matching score;

[0167] Based on the literature co-published by the target author and the author to be analyzed, indirect ID association between the target author and the author to be analyzed is established to obtain a third matching score;

[0168] Based on the source, verification state, and frequency of use of the identity recognition code, a credibility weight is configured for the first matching score, the second matching score, and the third matching score, respectively, and based on the configured credibility weight and the three matching scores, a result of the identity recognition code similarity mining of the target author and the author to be analyzed is generated.

[0169] As introduced above, the enhanced ID matching promotion is used in the embodiments of the present application, which considers three levels of ID matching, i.e., the matching of ORCID, the matching of Scopus ID, and the matching of indirect ID association.

[0170] When the ORCID complete match is performed, if the ORCID complete match is reached, a first matching score 1.0 can be obtained. The fuzzy match of the Scopus ID can tolerate 1-2 character differences. The published works or literatures each have a corresponding literature ID. In the embodiment of the present application, based on the literature co-published by the target author and the author to be analyzed, an indirect ID association between the target author and the author to be analyzed is established, and a third matching score is obtained. Specifically, the ID of the co-published literature is used to construct the indirect ID association between the target author and the author to be analyzed.

[0171] When the identity recognition code is mined, not only the ID matching itself is considered, but also the evaluation of the ID credibility is considered. As introduced above, for each matching score, the credibility weight can be configured according to the source, verification state and usage frequency of the identity recognition code. Thus, the final identity recognition code similarity mining result has higher credibility, and the accuracy of author name disambiguation is improved.

[0172] The above describes the similarity mining process of the six dimensions of identity recognition code, name, institution, cooperation network, time and alias in detail. In the foregoing embodiment, it is introduced in step S103 that the results of the similarity mining of the multiple dimensions between the same author information in the candidate information set and the author information to be analyzed are fused to obtain a multiple dimension similarity fusion result. In the following, the example adjustment scheme of the weight of the fusion process is introduced in combination with the mining of the six dimensions. In the embodiment of the present application, it is mentioned that the weight used for the fusion of each dimension can be adaptively adjusted based on the data quality.

[0173] First, the basic weight corresponding to each of the six dimensions of identity recognition code, name, institution, cooperation network, time and alias is obtained. The sum of the basic weights corresponding to the six dimensions is 1. As an example, the basic weights corresponding to the six dimensions of identity recognition code, name, institution, cooperation network, time and alias are respectively: 0.25 (identity recognition code), 0.4 (name), 0.15 (institution), 0.15 (cooperation network), 0.03 (time) and 0.02 (alias).

[0174] If the author information meets the data quality-based weight adaptive adjustment condition or the author information meets the scene-based weight adaptive adjustment condition, the basic weight corresponding to each of the six dimensions is dynamically adjusted according to the weight dynamic adjustment scheme corresponding to the condition met.

[0175] The data quality-based weight adaptive adjustment condition described above includes one or more of the following conditions: a first condition, a second condition and a third condition.

[0176] The first condition is that the credibility of the identity recognition code information in the author information is greater than a preset credibility threshold.

[0177] The second condition is that the name information in the author information is incomplete.

[0178] The third condition is that the organization information in the author information is missing.

[0179] The scenario-based weight self-adaptive adjustment condition includes one or more of the fourth condition, the fifth condition, and the sixth condition.

[0180] The fourth condition is that the author information meets an international cooperation scenario.

[0181] The fifth condition is that the author information meets a cross-disciplinary research scenario.

[0182] The sixth condition is that the author information is identified as an emerging scholar.

[0183] In the embodiments of the present application, it is considered that any one of the first condition to the sixth condition is met, and the basic weight of each corresponding dimension is dynamically adjusted according to the dynamic adjustment scheme of the met condition. The reason for adjusting the basic weight of the six dimensions is that the sum of the basic weights of the six dimensions is 1, and if one weight is adjusted up or down, the basic weight of the remaining dimensions also changes to meet the requirement that the sum of the basic weights is 1.

[0184] The weight adjustment scheme corresponding to the first condition is to increase the weight corresponding to the identity recognition code and proportionally reduce the weights corresponding to the other dimensions. As an example, when the credibility of the identity recognition code dimension is high, the weight of the identity recognition code is increased from 0.25 to 0.4, and the basic weights of the other dimensions are proportionally reduced.

[0185] The weight adjustment scheme corresponding to the second condition is to increase the weight corresponding to the cooperation network and proportionally reduce the weights corresponding to the other dimensions. As an example, when the name dimension data is incomplete, the weight of the cooperation network dimension is increased to 0.25, and the weights corresponding to the other dimensions are proportionally reduced.

[0186] The weight adjustment scheme corresponding to the third condition is to increase the sum of the weights of the time and alias dimensions and proportionally reduce the weights corresponding to the other dimensions. When the organization information is missing, the sum of the weights of the time and alias dimensions is increased to 0.1, and the weights corresponding to the other dimensions are proportionally reduced. Thus, the proportion of the similarity mining result of the time and alias dimensions in the fused mining result is increased.

[0187] The weight adjustment scheme corresponding to the fourth condition is to increase the weight corresponding to the organization and proportionally reduce the weights corresponding to the other dimensions. For example, the weight corresponding to the organization is further increased by 0.05, and the weights corresponding to the other dimensions are proportionally reduced.

[0188] The weight adjustment scheme corresponding to the fifth condition is to increase the weight corresponding to the cooperative network and proportionally reduce the weights corresponding to other dimensions. For example, the weight corresponding to the cooperative network is further increased by 0.1, and the weights corresponding to other dimensions are proportionally reduced.

[0189] The weight adjustment scheme corresponding to the sixth condition is to increase the weight of time and proportionally reduce the weights of other dimensions. For example, the weight of the time dimension is further increased by 0.05, and the weights of other dimensions are proportionally reduced.

[0190] After the adaptive adjustment of the weights described above, the implementation process of step S103 can be specifically as follows: the results of the identity recognition code similarity mining, the name similarity mining, the organization similarity mining, the cooperative network similarity mining, the time similarity mining and the alias similarity mining between the same author information in the candidate information set and the author information to be analyzed are fused according to the adjusted weights of the corresponding dimensions.

[0191] In the present application, through the innovative dynamic weight adjustment mechanism, on the one hand, the data quality of each dimension (such as integrity, reliability, consistency and other factors reflecting data quality) is considered, and on the other hand, international cooperation, interdisciplinary research and emerging scholars and other scene elements are fully considered. Therefore, the fusion process of the similarity mining results of multiple dimensions is matched with the scene and fully considers the data quality, achieving a dynamic and flexible fusion effect.

[0192] For the evaluation of data quality, the integrity, reliability and consistency can all be scored. For example, for the score of integrity, the ratio of the number of available data dimensions to the total number of data dimensions is used as the score result. The reliability can be evaluated according to the data source, verification status and historical accuracy. The consistency score is to verify the consistency of information between multiple dimensions.

[0193] In addition to the dynamic adjustment and optimization of the weights of the dimensions introduced above, the present application embodiment also proposes to apply a multi-level rule engine in the stage of fusing the similarity mining results of the dimensions. The multi-level rule engine helps to determine whether the fusion result needs to be corrected, and thus effectively improves the accuracy of author name disambiguation. The application of the multi-level rule engine will be introduced below.

[0194] In the embodiments of the present application, the results of identity recognition code similarity mining, the results of name similarity mining, the results of institution similarity mining, the results of cooperation network similarity mining, the results of time similarity mining and the results of alias similarity mining between the same author information and the author information to be analyzed in the candidate information set are fused according to the weights of the corresponding dimensions to obtain a plurality of dimension similarity fusion results between the same author information and the author information to be analyzed, including:

[0195] The results of identity recognition code similarity mining, the results of name similarity mining, the results of institution similarity mining, the results of cooperation network similarity mining, the results of time similarity mining and the results of alias similarity mining between the same author information and the author information to be analyzed in the candidate information set are fused according to the weights of the corresponding dimensions to obtain a preliminary fusion result.

[0196] Based on the multi-level rule engine, it is determined whether the preliminary fusion result needs to be corrected. If it is determined that the preliminary fusion result needs to be corrected, the correction result of the preliminary fusion result is obtained based on the multi-level rule engine, and the correction result is taken as the plurality of dimension similarity fusion results. If it is determined that the preliminary fusion result does not need to be corrected, the preliminary fusion result is taken as the plurality of dimension similarity fusion results.

[0197] The multi-level rule engine includes at least two categories of rules:

[0198] The rules include a forced matching rule, an inhibition rule, an enhancement rule or a risk identification rule.

[0199] The forced matching rule includes at least one of a first rule and a second rule.

[0200] The first rule is: if the identity recognition codes are completely matched, the lower limit of the plurality of dimension similarity fusion results is controlled. For example, if the ORCID is completely matched, the overall similarity (i.e. the plurality of dimension similarity fusion results) is guaranteed to reach 0.95 at the lowest.

[0201] The second rule is: if the name is the same, the institution is the same and there are more than three common collaborators, the lower limit of the plurality of dimension similarity fusion results is controlled. For example, under the second rule, the overall similarity is guaranteed to reach 0.90 at the lowest.

[0202] The inhibition rule includes at least one of a third rule, a fourth rule and a fifth rule.

[0203] The third rule is: if the result of name similarity mining is less than a first preset value, the upper limit of the plurality of dimension similarity fusion results is controlled. For example, the first preset value is 0.2. If the similarity mining result of the name dimension is <0.2, the overall similarity is required to be no more than 0.3.

[0204] The fourth rule is to control the upper limit of the fusion result of multiple dimensions if the institutional country does not match and there is no international cooperation history. For example, under the fourth rule, the overall similarity is required to be less than 0.4.

[0205] The fifth rule is to control the upper limit of the fusion result of multiple dimensions if the publication year difference exceeds 15 years and there is no continuous cooperation. For example, under the fifth rule, the overall similarity is required to be less than 0.5.

[0206] The enhancement rules include at least one of the sixth rule, the seventh rule and the eighth rule.

[0207] The sixth rule is to increase the value on the basis of the preliminary fusion result if the similarity mining result of three or more dimensions all exceeds the second preset value. As an example, the second preset value is 0.8, and under the sixth rule, the overall similarity is increased by 0.15 on the basis of the preliminary fusion result.

[0208] The seventh rule is to increase the value on the basis of the preliminary fusion result if the network centrality similarity of the cooperation network exceeds the third preset value and the institutional matching. For example, under the seventh rule, the overall similarity is increased by 0.1 on the basis of the preliminary fusion result.

[0209] The eighth rule is to increase the value on the basis of the preliminary fusion result if the publication mode similarity exceeds the fourth preset value. The publication mode is a behavior mode formed by multiple time dimensions including publication frequency, publication active period distribution, career trajectory evolution and annual output rhythm. The publication mode similarity exceeding the fourth preset value indicates that the publication mode is highly similar. As an example, under the eighth rule, the overall similarity is increased by 0.08 on the basis of the preliminary fusion result.

[0210] The risk identification rule includes at least one of the following:

[0211] If an abnormally high score is detected in the preliminary fusion result, a secondary verification mechanism is automatically triggered to check whether there is a data quality problem and to identify whether there is a malicious matching behavior;

[0212] Based on the checked data quality problem or the identified malicious matching behavior, at least one of the following measures is performed:

[0213] Lower the confidence of the corresponding match, mark it for manual review or reject the match.

[0214] As an example, in the risk identification rule, by setting a similarity threshold upper limit (such as 0.95 or above), abnormal situations are detected: when the preliminary fusion result shows an abnormally high score, the system will automatically trigger a secondary verification mechanism to check whether there are data entry errors (such as duplicate records, abnormal formats), identity code conflicts (the same ORCID is used by multiple authors), tampering of institutional information, and other data quality problems, while identifying malicious matching behaviors (such as deliberately forging cooperative relationships, batch false identity association), once such risks are detected, the system will reduce the confidence of the matching, mark it for manual review or directly reject the matching, to ensure the reliability and security of the disambiguation result.

[0215] As can be known from the above description, the embodiments of the present application can detect abnormal situations with high similarity by using risk identification rules and confirm them with manual review, thereby eliminating the interference of abnormally high scores on the accuracy of author name disambiguation. In addition, by identifying data quality problems and discovering possible malicious matching attempts, the safety and reliability of the entire author name disambiguation process are further ensured.

[0216] In the stage of fusing the similarity mining results in multiple dimensions, parallel optimal matching search is also proposed. For example, batch parallel processing is implemented: the candidate information set is calculated in parallel in batches of 15; thread pool optimization strategy is adopted: dynamically adjust the number of threads (2-6), avoid resource competition, memory-friendly batch processing: control memory usage to prevent memory overflow (Out Of Memory, OOM) caused by large data sets.

[0217] In addition, intelligent result mapping and verification are also performed on the fusion results of multiple dimension similarities. Specifically, it includes: entity ID unified mapping: consistency mapping of the alias table author_id to the entity table data_id; result confidence evaluation: based on similarity distribution, historical success rate, and comprehensive data quality score; abnormal result detection: identify obviously unreasonable matching results and trigger the manual review process.

[0218] The embodiments of the present application can use the intelligent double-stage adaptive cache introduced above to initialize the system. In addition, the embodiments of the present application also provide a three-stage cache optimization system. In the architecture design of the cache, L1 cache (memory LRU), L2 cache (SQLite persistence) and L3 cache (distributed cache) are included. Among them, the L1 cache capacity is 3000, which is adaptively expanded to 10000; TTL: dynamically adjusted based on access frequency (1-24 hours); hit rate monitoring: target 95% or more. The L2 cache capacity is 1 million, which is regularly cleaned up expired data; index optimization: composite index, query time <5ms; concurrent control: WAL mode, supporting multiple reading and single writing. The L3 cache is configured with Redis cluster, supporting TB-level data storage; consistent hashing, supporting horizontal expansion; data sharding, avoiding hot spot problems.

[0219] In addition, the embodiments of the present application propose a distributed computing framework, a task allocation strategy based on candidate information set size and computing complexity dynamic sharding; polling + weighting strategy, a load balancing algorithm for distributing tasks according to node performance; and a fault tolerance recovery mechanism supporting automatic retry and breakpoint continuation of tasks.

[0220] In the embodiments of the present application, the database connection pool is also optimized. It is proposed that thread local storage, each thread is independently connected, so as to avoid lock competition. It is proposed that connection reuse strategy, keep-alive of idle connection, dynamic expansion in busy time. It is proposed that connection health check, periodic detection of connection state, automatic reconnection.

[0221] In addition, the embodiments of the present application also propose performance monitoring and tuning measures. Real-time performance indicators such as QPS, delay distribution, error rate, and resource usage are analyzed, and an automatic tuning mechanism that automatically adjusts parameters according to load is used. In addition, a warning system is used to alarm and automatically recover in time when the system performance is abnormal.

[0222] From the above, the embodiments of the present application propose a high-performance optimization architecture with a distributed parallel computing framework, a three-stage cache optimization system, a database connection pool optimization, and a performance monitoring and tuning mechanism. The use of this system architecture enables the author name disambiguation method proposed in the present application to be calculated in a more efficient manner, and the query and result storage mode and strategy are optimized. The cooperation of the stable and reliable system and the real-time monitoring mechanism and alarm mechanism improves the user's sensitivity to system abnormalities.

[0223] Figure 5 An implementation architecture diagram of an author name disambiguation method provided by the embodiments of the present application. As shown in Figure 5 , the implementation architecture involves five stages:

[0224] First stage: intelligent initialization and network construction.

[0225] Second stage: construction and filtering of intelligent candidate information set.

[0226] Third stage: multi-dimensional similarity calculation of innovation.

[0227] Fourth stage: adaptive decision fusion mechanism.

[0228] Fifth stage: high-performance optimization architecture.

[0229] In the first stage, it involves the initialization of a two-level cache system, the construction of a multi-level cooperative network, the construction of an institutional hierarchical identification system, a standardized mapping table of country codes, and a library of institutional geographic location and subject distribution characteristics. In addition, it also implements an alias clustering algorithm.

[0230] In the second stage, it involves multi-dimensional input analysis of author information, hierarchical filtering strategy, and optimization of intelligent query method.

[0231] In the third stage, it involves enhanced identity recognition code matching algorithm, deep name similarity calculation, institutional intelligent matching algorithm, deep analysis of the topology of cooperative network, intelligent analysis of time dimension, and intelligent expansion matching of aliases.

[0232] In the fourth stage, it involves innovative dynamic weight optimization method fusion mechanism, multi-level rule engine, parallel optimal matching search, and mapping and verification of matching results.

[0233] In the fifth stage, it involves distributed parallel computing framework, three-level cache optimization system, database connection pool optimization, and performance monitoring and tuning.

[0234] Based on the analysis and introduction above, the existing technology has the following problems:

[0235] (1) The problem of insufficient depth of network analysis:

[0236] Although the existing technology considers cooperative networks, it only stays at the level of simple co-author counting and fails to explore the topological structure characteristics (such as degree centrality, clustering coefficient, and shortest path) and dynamic evolution patterns of the network, resulting in limited recognition ability of complex academic relationships.

[0237] (2) The problem of static similarity fusion mechanism:

[0238] The existing method uses fixed weights for multi-dimensional fusion, which cannot dynamically adjust according to data quality differences, dimension reliability, and specific application scenarios, and significantly decreases in effectiveness when dealing with incomplete data or special scenarios.

[0239] (3) The problem of original calculation efficiency optimization strategy:

[0240] The prior art lacks candidate information set pre-filtering algorithms and hierarchical cache optimization mechanisms. When facing million-level author data, the computational complexity increases exponentially, which cannot meet the real-time application requirements.

[0241] (4) Roughness of name standardization processing:

[0242] The existing method is insufficient in recognizing composite surnames, standardizing Unicode characters, and unifying the format of multi-language names, and the accuracy rate decreases significantly when processing international academic data.

[0243] (5) Simplification of organization matching algorithm:

[0244] The prior art does not establish an organization level recognition system and a standardized mapping of country codes, and cannot effectively handle the diversification of organization names and hierarchical relationships.

[0245] The present application proposes innovative technical paths to solve the above problems:

[0246] (1) Deep network structure analysis technology:

[0247] A multi-level academic cooperation network model is constructed, and the topological features such as degree centrality, clustering coefficient, and betweenness centrality of nodes are calculated by graph neural network algorithm to realize deep similarity analysis based on network structure.

[0248] (2) Self-adaptive multi-dimensional fusion algorithm:

[0249] A data quality evaluation model and dynamic weight adjustment mechanism are designed to adjust the fusion weight in real time according to the integrity, credibility, and historical matching success rate of each dimension of data, and to improve the overall matching accuracy.

[0250] (3) Hierarchical intelligent cache optimization system:

[0251] A three-level cache architecture of memory LRU cache + SQLite persistent cache + distributed cache is created, combined with query prediction algorithm and hot data identification mechanism to realize millisecond-level response.

[0252] (4) Multi-language intelligent standardization engine:

[0253] A name standardization system is established, which includes composite surname recognition algorithm, Unicode standardization processing, and multi-language format unification, supporting mixed processing of Chinese and English and other languages.

[0254] The core technology highlights in the technical solution of the present application are as follows:

[0255] 1. Original network topology depth analysis algorithm: Unlike the simple collaborator count of the prior art, the present application realizes network structure similarity calculation based on graph neural network, including advanced network features such as degree centrality, clustering coefficient, and structural hole analysis, filling the technical gap of academic network depth analysis;

[0256] 2. Breakthrough adaptive weight fusion mechanism: Innovative design of data quality evaluation model and dynamic weight adjustment algorithm, real-time optimization of fusion weight according to the integrity, reliability and historical success rate of each dimension data, better than the fixed weight method;

[0257] 3. Original three-level intelligent cache architecture: The three-level cache system of memory LRU + SQLite persistence + distributed cache is combined with query prediction and hotspot identification to achieve 99.5% cache hit rate and millisecond-level response speed;

[0258] 4. First multi-language intelligent standardization engine: Establish a complete system of composite surname recognition, Unicode standardization and multi-language format unification, support mixed processing of Chinese and English and standardization of institutions in 200+ countries / regions;

[0259] 5. Innovative multi-level special rule engine: Design four types of intelligent decision rules including mandatory matching, suppression rules, enhancement rules and risk identification to achieve accurate discrimination and abnormal identification in complex scenarios;

[0260] 6. Unique distributed parallel computing framework: Implement a complete distributed architecture of task fragmentation, load balancing and fault recovery to support real-time processing of million-level author data, with significantly improved computing efficiency compared to traditional methods;

[0261] 7. Original network evolution time series analysis: Time series network analysis is introduced for the first time in author disambiguation to identify author identity through the time evolution pattern of cooperative relationship, solving the problem of ignoring time dynamics in traditional methods.

[0262] Based on the above embodiment, the present application also proposes an author name disambiguation device. Figure 6 The structure diagram of the author name disambiguation device is shown in Figure 6 The device includes:

[0263] An information acquisition module 61 is configured to acquire a piece of author information to be analyzed and a candidate information set; the candidate information set includes multiple pieces of author information; and the author information to be analyzed includes a collaborator list.

[0264] The multi-dimension similarity mining module 62 is configured to perform multi-dimension similarity mining between the plurality of author information and the author information to be analyzed respectively; the cooperation network is one of the plurality of dimensions, the authors are represented by nodes in the cooperation network, and the cooperation relationship between the authors is represented by a connection between the nodes; the similarity mining of the cooperation network includes at least one of direct cooperation relationship similarity mining based on the cooperation network, indirect network relationship similarity mining based on the cooperation network, network structure topology feature similarity mining based on the cooperation network, or time sequence network evolution similarity mining based on the cooperation network;

[0265] The multi-dimension similarity fusion module 63 is configured to perform fusion processing on the results of the multi-dimension similarity mining between the same author information in the candidate information set and the author information to be analyzed, to obtain multi-dimension similarity fusion results.

[0266] The matching module 64 is configured to obtain a matching result of the author to be analyzed and the author in the candidate information set according to the multi-dimension similarity fusion results of the plurality of author information and the author information to be analyzed respectively.

[0267] In an optional implementation, the multi-dimension similarity mining module 62 is specifically configured to:

[0268] analyze the network structure topology feature of the first cooperation network to obtain a local clustering coefficient, a structural hole position and a network centrality of the first cooperation network; the first cooperation network is a cooperation network about the author to be analyzed;

[0269] analyze the network structure topology feature of the second cooperation network to obtain a local clustering coefficient, a structural hole position and a network centrality of the second cooperation network; the second cooperation network is a cooperation network about the target author;

[0270] calculate a local clustering similarity between the target author and the author to be analyzed based on the local clustering coefficient of the first cooperation network and the local clustering coefficient of the second cooperation network;

[0271] calculate a structural hole position similarity between the target author and the author to be analyzed based on the structural hole position of the first cooperation network and the structural hole position of the second cooperation network;

[0272] calculate a network centrality similarity between the target author and the author to be analyzed based on the network centrality of the first cooperation network and the network centrality of the second cooperation network;

[0273] According to the local clustering similarity, the structural hole position similarity and the network centrality similarity, a network structure topology feature similarity mining result of the target author and the author to be analyzed is calculated.

[0274] In an optional implementation, the multi-dimension similarity mining module 62 is specifically configured to:

[0275] perform time window analysis based on a first cooperation network to obtain a first time window analysis result and identify a dynamic change pattern of the first cooperation network; the first cooperation network is a cooperation network about the author to be analyzed;

[0276] perform time window analysis based on a second cooperation network to obtain a second time window analysis result and identify a dynamic change pattern of the second cooperation network; the second cooperation network is a cooperation network about the target author;

[0277] obtain a network time window similarity of the target author and the author to be analyzed based on the first time window analysis result and the second time window analysis result;

[0278] obtain a network dynamic change pattern similarity of the target author and the author to be analyzed based on the dynamic change pattern of the first cooperation network and the dynamic change pattern of the second cooperation network;

[0279] According to the network time window similarity and the network dynamic change pattern similarity, a time sequence network evolution similarity mining result of the target author and the author to be analyzed is calculated.

[0280] In an optional implementation, the multi-dimension similarity mining module 62 is specifically configured to:

[0281] According to a first cooperation network, a first cooperation strength and a first cooperation density of the author to be analyzed and each author in the coauthor list are calculated; the first cooperation network is a cooperation network about the author to be analyzed;

[0282] According to a second cooperation network, a second cooperation strength and a second cooperation density of the target author and each author in the coauthor list are calculated; the second cooperation network is a cooperation network about the target author;

[0283] According to the first cooperation strength and the second cooperation strength of the author to be analyzed and the target author with the same author in the coauthor list, a cooperation strength similarity of the target author and the author to be analyzed is obtained;

[0284] According to the first cooperation density and the second cooperation density of the same author in the coauthor list of the target author and the analyzed author, a cooperation density similarity between the target author and the analyzed author is obtained;

[0285] According to the cooperation intensity similarity and the cooperation density similarity, a direct cooperation relationship similarity mining result between the target author and the analyzed author is obtained.

[0286] In an optional implementation, the multi-dimension similarity mining module 62 is specifically configured to:

[0287] Identify a two-degree connection node and a three-degree connection node of the analyzed author from a first cooperation network; the first cooperation network is a cooperation network about the analyzed author;

[0288] Identify a two-degree connection node and a three-degree connection node of the target author from a second cooperation network; the second cooperation network is a cooperation network about the target author;

[0289] Obtain a two-degree connection similarity between the target author and the analyzed author based on the two-degree connection node of the analyzed author and the two-degree connection node of the target author;

[0290] Obtain a three-degree connection similarity between the target author and the analyzed author based on the three-degree connection node of the analyzed author and the three-degree connection node of the target author;

[0291] According to the two-degree connection similarity and the three-degree connection similarity, an indirect network relationship similarity mining result between the target author and the analyzed author is obtained.

[0292] In an optional implementation, the similarity mining of the cooperation network includes: direct cooperation relationship similarity mining based on the cooperation network, indirect network relationship similarity mining based on the cooperation network, network structure topology feature similarity mining based on the cooperation network, and time sequence network evolution similarity mining based on the cooperation network;

[0293] The multi-dimension similarity mining module 62 is further configured to: based on the network structure topology feature similarity mining result, the time sequence network evolution similarity mining result, the direct cooperation relationship similarity mining result, and the indirect network relationship similarity mining result, fuse to obtain a result of the cooperation network similarity mining.

[0294] In an optional implementation, the multiple dimensions further include: an identity code, a name, an institution, a time, and an alias;

[0295] The multi-dimension similarity fusion module 63 is specifically configured to:

[0296] The results of the identity code similarity mining, the name similarity mining, the organization similarity mining, the cooperation network similarity mining, the time similarity mining and the alias similarity mining between the same author information and the author information to be analyzed in the candidate information set are fused according to the weights of the corresponding dimensions to obtain a plurality of dimension similarity fusion results between the same author information and the author information to be analyzed.

[0297] In an optional implementation, the author name disambiguation apparatus further includes:

[0298] The basic weight acquisition module is configured to acquire a basic weight corresponding to each of the six dimensions of identity code, name, organization, cooperation network, time and alias, and the sum of the basic weights corresponding to the six dimensions is 1;

[0299] The weight self-adaptive adjustment module is configured to, if the author information satisfies a data quality-based weight self-adaptive adjustment condition or the author information satisfies a scenario-based weight self-adaptive adjustment condition, dynamically adjust the basic weight corresponding to each of the six dimensions according to a weight dynamic adjustment scheme corresponding to the satisfied condition.

[0300] The multi-dimension similarity fusion module 63 is specifically configured to fuse the results of the identity code similarity mining, the name similarity mining, the organization similarity mining, the cooperation network similarity mining, the time similarity mining and the alias similarity mining between the same author information and the author information to be analyzed in the candidate information set according to the adjusted weights of the corresponding dimensions.

[0301] In an optional implementation, the multi-dimension similarity fusion module 63 is specifically configured to:

[0302] fuse the results of the identity code similarity mining, the name similarity mining, the organization similarity mining, the cooperation network similarity mining, the time similarity mining and the alias similarity mining between the same author information and the author information to be analyzed in the candidate information set according to the weights of the corresponding dimensions to obtain a preliminary fusion result;

[0303] determine whether the preliminary fusion result needs to be corrected based on a multi-level rule engine, if it is determined that the preliminary fusion result needs to be corrected, obtain a correction result of the preliminary fusion result based on the multi-level rule engine, and take the correction result as the multi-dimension similarity fusion result, and if it is determined that the preliminary fusion result does not need to be corrected, take the preliminary fusion result as the multi-dimension similarity fusion result;

[0304] The multi-level rule engine includes at least two categories of rules:

[0305] The mandatory matching rule, the suppression rule, the enhancement rule, or the risk identification rule.

[0306] In an optional implementation, the multi-dimensional similarity mining module 62 is configured to:

[0307] standardizing and preprocessing the institution information in the target author information and the institution information in the author information to be analyzed respectively, wherein the standardizing and preprocessing includes removing prefixes of institution types, extracting core keywords of institutions, and applying country codes for standardized mapping;

[0308] calculating Jaccard similarity and fuzzy matching similarity for the keywords of the two sets of institution information after the standardizing and preprocessing;

[0309] calculating an initial institution similarity based on the Jaccard similarity and the fuzzy matching similarity;

[0310] optimizing the initial institution similarity based on matching of geographical positions and matching of institution levels of the two sets of institution information, and taking the optimized similarity result as a result of institution similarity mining of the target author and the author to be analyzed.

[0311] In an optional implementation, the author name disambiguation apparatus further includes a construction module configured to:

[0312] constructing an institution level identification system, wherein the level of a university is higher than the level of a research institute, and the level of the research institute is higher than the level of a department;

[0313] constructing a country code standardized mapping table, wherein the country code standardized mapping table supports more than 200 countries or regions in multiple expression forms;

[0314] constructing an institution geographical position and subject distribution feature library;

[0315] the applying country codes for standardized mapping specifically includes mapping country codes in the institution information based on the country code standardized mapping table;

[0316] the matching of geographical positions of the two sets of institution information is determined based on analysis of the institution geographical position and subject distribution feature library;

[0317] the matching of institution levels of the two sets of institution information is determined based on analysis of the institution level identification system.

[0318] In an optional implementation, the multi-dimensional similarity mining module 62 is configured to:

[0319] respectively perform Unicode standardization processing on the name information in the target author information and the name information in the author information to be analyzed;

[0320] perform compound surname recognition on the two sets of name information after the Unicode standardization processing;

[0321] based on the complete matching weight of the surname, the fuzzy matching weight of the surname, the full name Token sorting matching weight, and the initial sequence matching weight, calculate a multi-level matching result of the two sets of name information;

[0322] perform cultural background matching on the two sets of name information; wherein the cultural background matching includes: for Chinese names, perform pinyin conversion and initial matching;

[0323] generate a name similarity mining result of the target author and the author to be analyzed according to the multi-level matching result and the result of the cultural background matching.

[0324] In an optional implementation, the multi-dimensional similarity mining module 62 is configured to:

[0325] calculate a publication year similarity using a Gaussian decay function according to the publication year in the target author information and the publication year in the author information to be analyzed;

[0326] infer an academic career stage of the target author according to the first publication year in the target author information, and infer an academic career stage of the author to be analyzed according to the first publication year in the author information to be analyzed, perform career trajectory analysis based on the two sets of academic career stages, and obtain a career trajectory similarity analysis result;

[0327] calculate an overlap degree of the active period of works published according to the publication year in the target author information and the publication year in the author information to be analyzed;

[0328] calculate a publication frequency similarity of the target author and the author to be analyzed according to the annual publication quantity in the target author information and the annual publication quantity in the author information to be analyzed;

[0329] generate a time similarity mining result of the target author and the author to be analyzed according to the publication year similarity, the career trajectory similarity analysis result, the overlap degree of the active period of works published, and the publication frequency similarity.

[0330] In an optional implementation, the multi-dimensional similarity mining module 62 is configured to:

[0331] generate an alias for the name information in the target author information based on the author alias knowledge base, and generate an alias for the name information in the author information to be analyzed based on the author alias knowledge base;

[0332] perform an expansion match based on the two groups of generated aliases to obtain an alias match result;

[0333] perform a confidence evaluation for the alias match result based on the frequency of use, the time distribution, and the source credibility, and generate an alias similarity mining result of the target author and the author to be analyzed based on the evaluated confidence and the alias match result.

[0334] In an optional implementation, the multi-dimensional similarity mining module 62 is configured to:

[0335] perform a complete match for the ORCID in the target author information and the ORCID in the author information to be analyzed to obtain a first matching score;

[0336] perform a fuzzy match for the Scopus ID in the target author information and the Scopus ID in the author information to be analyzed using an edit distance algorithm to obtain a second matching score;

[0337] establish an indirect ID association between the target author and the author to be analyzed based on the literature co-published by the target author and the author to be analyzed to obtain a third matching score;

[0338] configure a credibility weight for the first matching score, the second matching score, and the third matching score based on the source, the verification state, and the frequency of use of the identity identifier, and generate a result of the identity identifier similarity mining of the target author and the author to be analyzed based on the configured credibility weight and the three matching scores.

[0339] In an optional implementation, the information acquisition module 61 is specifically configured to:

[0340] perform author information matching in the author information database based on the surname and the initial letter of the author information to be analyzed to obtain a first matching information set meeting a matching condition;

[0341] perform author information matching in the first matching information set based on the institutional keywords and the year range of the author information to be analyzed to obtain a second matching information set meeting a matching condition;

[0342] perform author information matching in the second matching information set based on the identity identifier of the author information to be analyzed to obtain a third matching information set meeting a matching condition as the candidate information set.

[0343] Based on the author name disambiguation method and device provided in the foregoing embodiments, the present application further provides an author name disambiguation device, which comprises a processor and a memory connected with each other in communication.

[0344] The memory stores a computer program.

[0345] The processor is configured to run the computer program to implement the author name disambiguation method of any implementation manner described in the method embodiments.

[0346] In addition, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is configured to implement the steps of the author name disambiguation method described in the method embodiments when executed by a processor.

[0347] It should be noted that each of the embodiments in the present specification is described in a progressive manner, and the same or similar parts of each embodiment can be referred to each other, and each embodiment mainly describes the difference from other embodiments. In particular, the device and equipment embodiments are basically similar to the method embodiments, so the description is relatively simple, and the relevant parts can be referred to the part of the method embodiment. The device and equipment embodiments described above are only illustrative, and the units described as separate components can be or can not be physically separated, and the components indicated as units can be or can not be physical units, that is, they can be located in one place, or can be distributed on multiple network units. According to the actual needs, part or all of the modules can be selected to achieve the purpose of the present embodiment. Those skilled in the art can understand and implement it without creative labor.

[0348] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any skilled person in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An author name disambiguation method characterized by, The method comprises the following steps: obtaining a piece of author information to be analyzed and a candidate information set; the candidate information set comprises a plurality of pieces of author information; the author information to be analyzed comprises a co-author list; respectively performing a plurality of dimension similarity mining between the plurality of pieces of author information and the author information to be analyzed; wherein the cooperation network is one of the plurality of dimensions, the authors are represented by nodes in the cooperation network, and the cooperation relationship between the nodes is represented by a connection between the nodes; the similarity mining of the cooperation network comprises at least one of the following: direct cooperation relationship similarity mining based on the cooperation network, indirect network relationship similarity mining based on the cooperation network, network structure topology feature similarity mining based on the cooperation network, or time sequence network evolution similarity mining based on the cooperation network; fusing the results of the plurality of dimension similarity mining between the same piece of author information in the candidate information set and the author information to be analyzed to obtain a plurality of dimension similarity fusion results; obtaining a matching result of the author to be analyzed and the authors in the candidate information set according to the plurality of dimension similarity fusion results of the plurality of pieces of author information and the author information to be analyzed respectively; the plurality of dimensions further comprise: identity code, name, institution, time and alias; the fusing the results of the plurality of dimension similarity mining between the same piece of author information in the candidate information set and the author information to be analyzed to obtain a plurality of dimension similarity fusion results comprises: fusing the results of the identity code similarity mining, the name similarity mining, the institution similarity mining, the cooperation network similarity mining, the time similarity mining and the alias similarity mining between the same piece of author information in the candidate information set and the author information to be analyzed according to the weights of the corresponding dimensions to obtain the plurality of dimension similarity fusion results between the same piece of author information and the author information to be analyzed; the method further comprises: obtaining the basic weights corresponding to the six dimensions of identity code, name, institution, cooperation network, time and alias respectively; the sum of the basic weights corresponding to the six dimensions is 1; if the author information meets the data quality-based weight self-adaptive adjustment condition or the author information meets the scene-based weight self-adaptive adjustment condition, dynamically adjusting the basic weights corresponding to the six dimensions according to the weight dynamic adjustment scheme corresponding to the met condition; the fusing the results of the identity code similarity mining, the name similarity mining, the institution similarity mining, the cooperation network similarity mining, the time similarity mining and the alias similarity mining between the same piece of author information in the candidate information set and the author information to be analyzed according to the weights of the corresponding dimensions, specifically comprises: The results of identity code similarity mining, name similarity mining, organization similarity mining, cooperation network similarity mining, time similarity mining and alias similarity mining between the same author information in the candidate information set and the author information to be analyzed are fused according to the adjusted weights of the corresponding dimensions.

2. The method of claim 1, wherein, The network structure topology feature similarity mining between the target author information in the candidate information set and the author information to be analyzed based on a cooperation network includes: analyzing the network structure topology features of a first cooperation network to obtain the local clustering coefficient, structural hole position and network centrality of the first cooperation network; the first cooperation network is the cooperation network about the author to be analyzed; analyzing the network structure topology features of a second cooperation network to obtain the local clustering coefficient, structural hole position and network centrality of the second cooperation network; the second cooperation network is the cooperation network about the target author; calculating the local clustering similarity between the target author and the author to be analyzed based on the local clustering coefficient of the first cooperation network and the local clustering coefficient of the second cooperation network; calculating the structural hole position similarity between the target author and the author to be analyzed based on the structural hole position of the first cooperation network and the structural hole position of the second cooperation network; calculating the network centrality similarity between the target author and the author to be analyzed based on the network centrality of the first cooperation network and the network centrality of the second cooperation network; calculating the network structure topology feature similarity mining result between the target author and the author to be analyzed according to the local clustering similarity, the structural hole position similarity and the network centrality similarity.

3. The method of claim 1, wherein, The time sequence network evolution similarity mining between the target author information in the candidate information set and the author information to be analyzed based on a cooperation network includes: performing time window analysis based on a first cooperation network to obtain a first time window analysis result and identify the dynamic change mode of the first cooperation network; the first cooperation network is the cooperation network about the author to be analyzed; performing time window analysis based on a second cooperation network to obtain a second time window analysis result and identify the dynamic change mode of the second cooperation network; the second cooperation network is the cooperation network about the target author; obtaining the network time window similarity between the target author and the author to be analyzed based on the first time window analysis result and the second time window analysis result; obtaining the network dynamic change mode similarity between the target author and the author to be analyzed based on the dynamic change mode of the first cooperation network and the dynamic change mode of the second cooperation network; calculating the time sequence network evolution similarity mining result between the target author and the author to be analyzed according to the network time window similarity and the network dynamic change mode similarity.

4. The method of claim 1, wherein, The target author information in the candidate information set and the author information to be analyzed are subjected to direct cooperation relationship similarity mining based on a cooperation network, including: According to a first cooperation network, first cooperation strength and first cooperation density of each author in the author list with respect to the author to be analyzed are calculated; the first cooperation network is a cooperation network with respect to the author to be analyzed; According to a second cooperation network, second cooperation strength and second cooperation density of each author in the author list with respect to the target author are calculated; the second cooperation network is a cooperation network with respect to the target author; According to the first cooperation strength and the second cooperation strength of the same author in the author list with respect to the author to be analyzed and the target author respectively, cooperation strength similarity of the target author with respect to the author to be analyzed is obtained; According to the first cooperation density and the second cooperation density of the same author in the author list with respect to the author to be analyzed and the target author respectively, cooperation density similarity of the target author with respect to the author to be analyzed is obtained; According to the cooperation strength similarity and the cooperation density similarity, direct cooperation relationship similarity mining result of the target author with respect to the author to be analyzed is calculated.

5. The method of claim 1, wherein, The target author information in the candidate information set and the author information to be analyzed are subjected to indirect network relationship similarity mining based on a cooperation network, including: Second-degree connection nodes and third-degree connection nodes of the author to be analyzed are identified from a first cooperation network; the first cooperation network is a cooperation network with respect to the author to be analyzed; Second-degree connection nodes and third-degree connection nodes of the target author are identified from a second cooperation network; the second cooperation network is a cooperation network with respect to the target author; Based on the second-degree connection nodes of the author to be analyzed and the second-degree connection nodes of the target author, second-degree connection similarity of the target author with respect to the author to be analyzed is obtained; Based on the third-degree connection nodes of the author to be analyzed and the third-degree connection nodes of the target author, third-degree connection similarity of the target author with respect to the author to be analyzed is obtained; According to the second-degree connection similarity and the third-degree connection similarity, indirect network relationship similarity mining result of the target author with respect to the author to be analyzed is calculated.

6. The method of claim 1, wherein, The similarity mining of the cooperation network includes: direct cooperation relationship similarity mining based on a cooperation network, indirect network relationship similarity mining based on a cooperation network, network structure topology feature similarity mining based on a cooperation network, and time sequence network evolution similarity mining based on a cooperation network; The method further includes: based on the network structure topology feature similarity mining result, the time sequence network evolution similarity mining result, the direct cooperation relationship similarity mining result and the indirect network relationship similarity mining result, a result of the cooperation network similarity mining is fused.

7. The method of claim 1, wherein, The condition for adaptive adjustment of the weight based on the data quality includes one or more of the first condition, the second condition and the third condition; The first condition is that the credibility of the identity recognition code information in the author information is greater than a preset credibility threshold; The second condition is that the name information in the author information is incomplete. The third condition is that the organization information in the author information is missing. The weight adjustment scheme corresponding to the first condition is to increase the weight corresponding to the identity identification code and proportionally reduce the weights corresponding to other dimensions. The weight adjustment scheme corresponding to the second condition is to increase the weight corresponding to the cooperation network and proportionally reduce the weights corresponding to other dimensions. The weight adjustment scheme corresponding to the third condition is to increase the sum of the weights of the time and alias dimensions and proportionally reduce the weights corresponding to other dimensions.

8. The method of claim 1, wherein, The scenario-based weight self-adaptive adjustment condition includes one or more conditions among a fourth condition, a fifth condition, and a sixth condition. The fourth condition is that the author information meets the international cooperation scenario. The fifth condition is that the author information meets the interdisciplinary research scenario. The sixth condition is that the author information is identified as an emerging scholar. The weight adjustment scheme corresponding to the fourth condition is to increase the weight corresponding to the organization and proportionally reduce the weights corresponding to other dimensions. The weight adjustment scheme corresponding to the fifth condition is to increase the weight corresponding to the cooperation network and proportionally reduce the weights corresponding to other dimensions. The weight adjustment scheme corresponding to the sixth condition is to increase the weight of the time and proportionally reduce the weights corresponding to other dimensions.

9. The method of claim 1, wherein, The results of the identity identification code similarity mining, the name similarity mining, the organization similarity mining, the cooperation network similarity mining, the time similarity mining, and the alias similarity mining between the same author information in the candidate information set and the author information to be analyzed are fused according to the weights of the corresponding dimensions to obtain a plurality of dimension similarity fusion results between the same author information and the author information to be analyzed, including: The results of the identity identification code similarity mining, the name similarity mining, the organization similarity mining, the cooperation network similarity mining, the time similarity mining, and the alias similarity mining between the same author information in the candidate information set and the author information to be analyzed are fused according to the weights of the corresponding dimensions to obtain a preliminary fusion result. Based on a multi-level rule engine, it is determined whether the preliminary fusion result needs to be corrected. If it is determined that correction is needed, the correction result of the preliminary fusion result is obtained based on the multi-level rule engine, and the correction result is taken as the plurality of dimension similarity fusion results. If it is determined that correction is not needed, the preliminary fusion result is taken as the plurality of dimension similarity fusion results. The multi-level rule engine includes at least two categories of rules: The forced matching rule, the suppression rule, the enhancement rule, or the risk identification rule.

10. The method of claim 9, wherein, The forced matching rule includes at least one of a first rule and a second rule. The first rule is that if the identity identification codes are completely matched, the lower limit of the plurality of dimension similarity fusion results is controlled. The second rule is that if the names are the same, the organizations are the same, and there are more than three common collaborators, the lower limit of the plurality of dimension similarity fusion results is controlled.

11. The method of claim 9, wherein, The inhibition rules include at least one of a third rule, a fourth rule and a fifth rule; The third rule is that if the result of the name similarity mining is less than a first preset value, the upper limit of the multiple-dimension similarity fusion result is controlled; The fourth rule is that if the institution country does not match and there is no international cooperation history, the upper limit of the multiple-dimension similarity fusion result is controlled; The fifth rule is that if the publication year difference exceeds 15 years and there is no continuous cooperation, the upper limit of the multiple-dimension similarity fusion result is controlled.

12. The method of claim 9, wherein, The enhancement rules include at least one of a sixth rule, a seventh rule and an eighth rule; The sixth rule is that if the results of similarity mining of three or more dimensions all exceed a second preset value, the value is adjusted upward on the basis of the preliminary fusion result; The seventh rule is that if the network centrality similarity of the cooperation network exceeds a third preset value and the institution matches, the value is adjusted upward on the basis of the preliminary fusion result; The eighth rule is that if the publication mode similarity exceeds a fourth preset value, the value is adjusted upward on the basis of the preliminary fusion result; the publication mode is a behavior mode formed by multiple time-dimension characteristics including publication frequency, publication active period distribution, career trajectory evolution and annual output rhythm.

13. The method of claim 9, wherein, The risk identification rules include at least one of the following: If it is detected that the preliminary fusion result has an abnormally high score, a secondary verification mechanism is automatically triggered to check whether there is a data quality problem and to identify whether there is a malicious matching behavior; Based on the checked data quality problem or the identified malicious matching behavior, at least one of the following measures is performed: The confidence of the corresponding match is reduced, marked for manual review or the match is rejected.

14. The method of claim 1, wherein, The institution-based similarity mining of target author information in the candidate information set and the author information to be analyzed includes: Standardized preprocessing is performed on the institution information in the target author information and the institution information in the author information to be analyzed; the standardized preprocessing includes removing the prefix of the institution type, extracting the core keywords of the institution and applying country code for standardized mapping; The Jaccard similarity and the fuzzy matching similarity of the keywords of the two sets of institution information after standardized preprocessing are calculated; An institution preliminary similarity is calculated based on the Jaccard similarity and the fuzzy matching similarity; Based on the matching of the geographical positions of the two sets of institution information and the matching of the institution levels, the institution preliminary similarity is optimized, and the optimized similarity result is taken as the result of the institution similarity mining of the target author and the author to be analyzed.

15. The method of claim 14, wherein, The method further includes: An institution level identification system is constructed; in the institution level identification system, the level of a university is higher than that of a research institute, and the level of the research institute is higher than that of a department; A country code standardized mapping table is constructed; the country code standardized mapping table supports more than 200 countries or regions in multiple expression forms; An institution geographical position and subject distribution feature library is constructed; The standardized mapping of the country code is specifically: the country code in the institution information is standardized mapped based on the country code standardized mapping table; The matching of the geographical positions of the two sets of institution information is determined based on analysis of the institution geographical position and subject distribution feature library; The matching of the institution levels of the two sets of institution information is determined based on analysis of the institution level identification system.

16. The method of claim 1, wherein, The target author information in the candidate information set and the author information to be analyzed are subjected to name-based similarity mining, including: The name information in the target author information and the name information in the author information to be analyzed are subjected to Unicode standardization processing respectively; Compound surnames in the two sets of name information subjected to Unicode standardization processing are identified; Based on the full-matching weight of surnames, the fuzzy-matching weight of surnames, the full-name Token sorting matching weight and the initial letter sequence matching weight, a multi-level matching result of the two sets of name information is calculated; Cultural background matching is adopted for the two sets of name information; wherein the cultural background matching includes: for Chinese names, pinyin conversion and initial letter matching are adopted; Based on the multi-level matching result and the result of cultural background matching, a name similarity mining result of the target author and the author to be analyzed is generated.

17. The method of claim 1, wherein, The target author information in the candidate information set and the author information to be analyzed are subjected to time-based similarity mining, including: Based on the publication years in the target author information and the publication years in the author information to be analyzed, a publication year similarity is calculated by using a Gaussian decay function; Based on the first publication year in the target author information, the academic career stage of the target author is inferred, and based on the first publication year in the author information to be analyzed, the academic career stage of the author to be analyzed is inferred, a career track analysis is performed based on the two sets of academic career stages, and a career track similarity analysis result is obtained; Based on the publication years in the target author information and the publication years in the author information to be analyzed, an overlap degree of work publication active periods is calculated; Based on the annual publication quantity in the target author information and the annual publication quantity in the author information to be analyzed, a publication frequency similarity of the target author and the author to be analyzed is calculated; Based on the publication year similarity, the career track similarity analysis result, the overlap degree of work publication active periods and the publication frequency similarity, a time similarity mining result of the target author and the author to be analyzed is generated.

18. The method of claim 1, wherein, The target author information in the candidate information set and the author information to be analyzed are subjected to alias-based similarity mining, including: Based on the author alias knowledge base, an alias is generated for the name information in the target author information, and based on the author alias knowledge base, an alias is generated for the name information in the author information to be analyzed; Based on the two sets of generated aliases, an extended matching is performed, and an alias matching result is obtained; Based on the usage frequency, time distribution and source credibility, a confidence evaluation is performed for the alias matching result, and based on the evaluated confidence and the alias matching result, an alias similarity mining result of the target author and the author to be analyzed is generated.

19. The method of claim 1, wherein, The target author information in the candidate information set and the author information to be analyzed are subjected to identity code-based similarity mining, including: For the ORCID in the target author information and the ORCID in the author information to be analyzed, complete matching is performed to obtain a first matching score; For the Scopus ID in the target author information and the Scopus ID in the author information to be analyzed, fuzzy matching is performed by using an edit distance algorithm to obtain a second matching score; Based on the literature co-published by the target author and the author to be analyzed, indirect ID association between the target author and the author to be analyzed is established to obtain a third matching score; Based on the source, verification state and use frequency of the identity code, a credibility weight is configured for the first matching score, the second matching score and the third matching score, and based on the configured credibility weight and the three matching scores, a result of identity code similarity mining between the target author and the author to be analyzed is generated.

20. The method of claim 1, wherein, The candidate information set is obtained through a three-level screening strategy based on the surname and initial letter, the institution keyword and the year range, and the identity code; obtaining the candidate information set includes: In the author information database, author information matching is performed based on the surname and initial letter of the author information to be analyzed to obtain a first matching information set meeting the matching condition; In the first matching information set, author information matching is performed based on the institution keyword and the year range of the author information to be analyzed to obtain a second matching information set meeting the matching condition; In the second matching information set, author information matching is performed based on the identity code of the author information to be analyzed to obtain a third matching information set meeting the matching condition as the candidate information set.

21. An author name disambiguation apparatus characterized by: It includes: An information acquisition module is configured to acquire a piece of author information to be analyzed and a candidate information set; The candidate information set includes multiple pieces of author information; The author information to be analyzed includes a co-author list; A multi-dimensional similarity mining module is configured to perform multi-dimensional similarity mining between the multiple pieces of author information and the author information to be analyzed, respectively; wherein the cooperation network is one of the multiple dimensions, the authors are represented by nodes in the cooperation network, and the cooperation relationship between the nodes is represented by a connection line; the similarity mining of the cooperation network includes at least one of the following: direct cooperation relationship similarity mining based on the cooperation network, indirect network relationship similarity mining based on the cooperation network, network structure topology feature similarity mining based on the cooperation network, or time sequence network evolution similarity mining based on the cooperation network; A multi-dimensional similarity fusion module is configured to perform fusion processing on the results of the multi-dimensional similarity mining between the same piece of author information in the candidate information set and the author information to be analyzed to obtain multi-dimensional similarity fusion results; A matching module is configured to obtain a matching result between the author to be analyzed and the authors in the candidate information set according to the multi-dimensional similarity fusion results of the multiple pieces of author information and the author information to be analyzed, respectively; The multiple dimensions further include: identity code, name, institution, time and alias. The multi-dimension similarity fusion module is specifically configured to: The results of the identity recognition code similarity mining, the results of the name similarity mining, the results of the organization similarity mining, the results of the cooperation network similarity mining, the results of the time similarity mining, and the results of the alias similarity mining between the same author information and the author information to be analyzed in the candidate information set are fused according to the weights of the corresponding dimensions to obtain multi-dimension similarity fusion results between the same author information and the author information to be analyzed. The device further includes: A basic weight acquisition module configured to acquire respective basic weights of six dimensions of identity recognition code, name, organization, cooperation network, time, and alias, wherein the sum of the respective basic weights of the six dimensions is 1. A weight self-adaptive adjustment module configured to, if the author information meets a weight self-adaptive adjustment condition based on data quality or the author information meets a weight self-adaptive adjustment condition based on a scenario, dynamically adjust the respective basic weights of the six dimensions according to a weight dynamic adjustment scheme corresponding to the met condition. The multi-dimension similarity fusion module is specifically configured to: fuse the results of the identity recognition code similarity mining, the results of the name similarity mining, the results of the organization similarity mining, the results of the cooperation network similarity mining, the results of the time similarity mining, and the results of the alias similarity mining between the same author information and the author information to be analyzed in the candidate information set according to the adjusted weights of the corresponding dimensions.

22. An author name disambiguation apparatus characterized by: It includes: A processor and a memory in communication with each other; The memory stores a computer program; The processor is configured to run the computer program to implement the author name disambiguation method according to any one of claims 1-20.

23. A computer-readable storage medium, characterized in that, The memory stores a computer program, and the computer program is configured to implement the steps of the author name disambiguation method according to any one of claims 1-20 when executed by the processor.

Images